From: John Ashburner 
Subject: Re: Help for constructing Template images
Date: Wed, 18 Dec 2002 16:22:59 +0000
To: SPM@JISCMAIL.AC.UK

> What are the advantages of customized template images
> in VBM analysis?

Customised templates are useful when:

1) The contrast in your MR images is not the same as the
   contrast used to generate the existing templates.  If
   the contrast is different, then the mean squared cost
   function is not optimal.  However, for "optimised" VBM
   this only really applies to the initial affine
   registration that is incorporated into the initial
   segmentation.  Contrast differences are likely to have
   a relatively small effect on the final results.

2) The demographics of your subject population differ
   from those used to generate the existing templates and
   prior probability images.  For example, serious problems
   can occur if your subjects have very large ventricles.
   In these data, there would be CSF in regions where the
   existing priors say CSF should not exist.  This would
   force some of the CSF to be classified as white matter,
   seriously affecting the intensity distribution that
   is used to model white matter.  This then has negative
   consequences for the whole of the segmentation.

> Can any one please explain the detailed steps to
> construct a customized template image (gray and white
> matter images) for VBM analysis?

The following script is one possible way of generating your
own template image.  Note that it takes a while to run, and
does not save any intermediate images that could be useful
for quality control.  Also, if it crashes at any point then
it is difficult to recover the work it has done so far.

* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 

make_template.m

* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 

You may also wish to do some manual editing of the images
afterwards - especially to remove extra-skull CSF.  When
everything has finished, simply smooth the images by 8mm
and call them templates and prior probability images.
You can modify the default priors for the segmentation step
in order that the customised ones are used.  This can be done
either by changing spm_defaults.m, or by typing the following
in Matlab:

spm_defaults
global defaults
defaults.segment.estimate.priors = ...
     spm_get(3,'*.IMAGE','Select GM,WM & CSF priors');

Note that this will be cleared if you reload the defaults.  This
could be done when you start spm, reset the defaults or if the
optimised VBM script is run, as it calls spm_defaults.m.
Alternatively the optimised VBM script could be modified to
include the above.

Note that I have only tried the script with three images, so
I don't have a good feel for how robust it is likely to be.

>
> Please let me know the number of subjects required to
> construct one?

Its hard to say, but more is best.  The 8mm smoothing means
that you can get away with slightly fewer than otherwise.

Best regards,
-John

Note: The make_template is an updated version of what was originally posted. It is current as of Sep 9, 2003.